Maligned #8 - Multimodal Gets Practical
Valentine’s Day, and I’m writing about AI. Make of that what you will.
Multimodal is finally useful, not just impressive
For the longest time, multimodal AI was the thing that looked great in demos and fell apart in production. That’s changing. The latest updates from Google, OpenAI, and Anthropic all show significant improvements in vision-language tasks that actually matter: document understanding, visual question answering over complex diagrams, and video analysis. I’m seeing real deployments now in insurance claims processing, manufacturing quality control, and medical imaging triage. The shift from “look what it can do” to “here’s how we’re using it” is the real story.
Open source AI licensing gets messy
The debate over what counts as “open source” in AI continues to be contentious. Meta’s Llama license, Mistral’s various license tiers, and the proliferation of “open-weight but not really open” releases have created genuine confusion. The Open Source Initiative has been trying to establish a clear definition, but the community is fractured. Some want strict copyleft-style licensing, others want permissive commercial terms, and the model providers want to call everything open source for the PR benefit while retaining control. This matters for enterprises making technology bets. You need to actually read the license, not just check the “open source” box.
AI in education is getting interesting pushback
Schools and universities spent the last year trying to either ban or embrace AI tools, with mixed results on both fronts. A new wave of thinking is emerging that’s more nuanced: redesigning assessments and curricula around the assumption that students will use AI, rather than fighting it. The early experiments are promising. Students who learn to use AI as a thinking tool, rather than a shortcut, seem to develop stronger analytical skills. But this requires teachers who understand the technology, and that’s a training gap nobody has closed yet.
Inference optimization is the unglamorous hero
While everyone argues about which model is best, a quieter revolution in inference optimization is doing more to make AI practical. Speculative decoding, KV-cache optimization, quantization improvements, and better batching strategies are collectively reducing inference costs and latency by margins that would have seemed impossible a year ago. This is the plumbing that makes everything else work, and the teams working on it deserve more recognition than they get.
See you next week.
Maligned - AI news by Mal